Boyun7/GroundProbe-ACT

GroundProbe ACT Checkpoints

Two Action Chunking with Transformers (ACT) policies trained on one cell of the GroundProbe demonstrations: quest_l2_cubes_pilot, instruction T1 ("Pick the left red block and place it in the bin") at the clean complexity, simulated Franka Panda in NVIDIA Isaac Lab-Arena.

ACT has no language input, so these are not grounding baselines. They exist to show that the demonstrations and the closed-loop evaluation path support learning the skill, and they are useful as a reference point when bringing up a new policy on the benchmark.

Code, scenes and the evaluation harness: https://github.com/AndersonYu7/Benchmark

Checkpoints

folder	action encoding	dataset column	qpos input	best epoch
delta_pose/	delta pose, 7-D	action	TCP pose, 7-D	510
joint/	absolute joint targets, 8-D	action.joint	7 joint positions + total finger opening	230

Each folder holds:

• policy_best.ckpt: the weights with the lowest validation loss
• config.json: architecture settings, the training/validation episode split and their layout seeds
• dataset_stats.pkl: the normalisation statistics the policy was trained with

config.json names the delta-pose encoding "action_space": "tcp"; that is the action column of the dataset.

Results

Closed-loop rollouts under the benchmark success criterion (target inside the bin for 30 consecutive control steps), at most 1000 control steps, executing the first 25 actions of every 50-action chunk before re-querying the policy.

policy	evaluation layouts	success	Wilson 95%
delta pose	the 20 training-demonstration layouts	17/20 (85%)	[64%, 95%]
joint	the 20 training-demonstration layouts	16/20 (80%)	[58%, 92%]
joint	the 5 held-out validation layouts	2/5 (40%)	[12%, 77%]
joint	20 layouts outside the collection seed range	5/20 (25%)	[11%, 47%]

With 50 demonstrations of one cell, the drop on unseen layouts is expected.

Training

ResNet-18 backbone, 4 encoder / 7 decoder layers, hidden size 512, feed-forward 3200, 8 heads, dropout 0.1, chunk size 50, both cameras (third_camera, wrist_camera) at 240×240, AdamW with learning rate 1e-5 (backbone 1e-5) and weight decay 1e-4, batch size 8, KL weight 10. 45 training and 5 validation episodes; validation every 10 epochs; stopped after 300 epochs without improvement.

Using them

The evaluation harness loads these through script/policy_adapters.py (ACTAdapter), which reads config.json and dataset_stats.pkl from the checkpoint's folder:

hf download Boyun7/GroundProbe-ACT --local-dir act_ckpt

python -m script.eval_policy --policy ACT --scene cubes \
    --ckpt act_ckpt/joint/policy_best.ckpt \
    --eval-episodes 20 --exec-horizon 25 --headless --out eval_out/act_joint \
    --environment l2_spatial_tasks.examples.manipulation.l2_spatial_env:L2SpatialEnv \
    --enable_cameras l2_spatial_env --embodiment franka --enable_cameras True

config.json records the dataset path as dataset/quest_l2_cubes_pilot, relative to the repository root, and the harness reads the control rate and the joint policy's starting configuration from it. Download the dataset into dataset/ at the repository root before evaluating.

License

Apache 2.0, matching the code and the dataset.